GPT-6 Astra: Released Flagship, API Pricing and Benchmarks

Astra is the GPT-6 flagship, not a Flash hop. Its $10 / $50 Standard list, 1.05M context, and reasoning.effort through max (the gateway default is low). OpenAI published a self-reported benchmark table. Hold those numbers as theirs, not as LLM Stats verified.

They call it their most capable model for the hardest end-to-end work: complex reasoning, coding, computer use, research, and document creation. That is a lab sentence. Hold the benchmark table, the sticker, and the released API status.

At a Glance

• Catalog id: gpt-6-astra
• Organization: OpenAI
• Status: API released September 4, 2026
• Model id: gpt-6-astra
• Pricing: $10 / $50 Standard; cache $1; writes $12.50
• Long prompts:>272K input → 2× input/cache, 1.5× output for the full request
• Context: 1,050,000 / 128K max output
• Cutoff: April 30, 2026
• Effort: low through max; gateway default low
• Modalities: text + image in, text out
• Hosting: first-party OpenAI in our catalog
• Self-reported: TB4 57.7 · DeepSWE 74.1 · TB-Science 64.6 · GPQA 96.0 · BrowseComp 91.5

What's New

GPT-6 family flagship. Context jumps to 1,050,000. reasoning.effort adds xhigh and max on top of low / medium / high. The Standard sticker stays $10 / $50 per 1M, with cached input at $1 and an explicit cache-write line at $12.50. The announcement carries a self-reported coding, science, computer-use, and cyber table.

Astra is now released through the OpenAI API as gpt-6-astra. The free tier is unsupported, and account access remains subject to OpenAI's rollout and rate limits. Path to Astra (September 1, 2026) designates Astra at the Critical cybersecurity capability threshold under OpenAI's Preparedness Framework, with advanced defender access limited through Trusted Access and Daybreak Blue.

What the card supports

Streaming, function calling, and structured outputs. Chat Completions and Responses are supported; Batch is supported; fine-tuning is not. Responses API tools include web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search. Input is text and image; output is text only.

Benchmarks

OpenAI published a self-reported benchmark table on the GPT-6 Astra announcement page. Numbers below are theirs, not LLM Stats verified. Read by job: coding, science, computer use, then cyber.

Coding

Terminal-Bench 4.0 lands at 57.7. DeepSWE v1.1 is 74.1, a small hop over GPT-5.6 Sol's 72.7 on OpenAI's own page. FrontierCode 1.1 Extended is 64.5; Main is 53.3.

Science and academic

Terminal-Bench-Science 0.1 at 64.6 vs Fable 5.1's 52.6 is the official comparison pair that matters on this page. FrontierMath T4 v2 is 97.6. HLE with tools is 57.2. GPQA Diamond at 96.0 is saturated; treat it as a ceiling row, not a mover.

Computer use and agents

BrowseComp is 91.5. OSWorld 2.0 is 72.6 offline partial. Agents' Last Exam is 59.3. AutomationBench is 41.4. ScreenSpot-Pro with no tools is 92.7; BenchCAD with Python is 95.9.

ARC

ARC-AGI-3 is 99.9, ARC-AGI-2 95.0, ARC-AGI-1 98.5. These are near-ceiling. The announcement page footnotes ARC-AGI-3 to a Responses API harness; read that row with the harness note attached.

Cyber and health

ExploitBench is 100%, a cyber eval of exploit development from known vulnerabilities, not the coding headline. ExploitGym is 42.4; SEC-Bench Pro is 85.4. Health rows on the same table: GeneBench Pro 37.8, LifeSciBench 60.3, HealthBench Professional (length-adjusted) 63.4.

Pricing

First-party OpenAI Standard, USD per million tokens. Same sticker class as recent frontier OpenAI lists. Compare to Claude Fable 5.1: same $10 / $50 list, different cache math (Fable hits $0.25 vs Astra's $1). OpenAI has not published a cache-hit workload percentage for Astra.

When to Use It

• Good fit: shops waiting for GPT-6 on hard coding, computer use, or long documents through the OpenAI API; 1.05M context jobs; teams that want xhigh or max effort.
• Watch the gateway default: low, not max. Set reasoning.effort explicitly.
• Prefer Fable 5.1 when you want Anthropic's published table and cheaper cache hits at matched $10 / $50 list.
• Cyber: the default product is not unrestricted dual-use. Advanced cyber workflows route through Trusted Access / Daybreak.
• Budget the 272K multiplier if prompts are huge. Fast is 2×.

Caveats

• Rollout and rate limits. The API model is released, but the free tier is unsupported and availability can still depend on the account and usage tier.
• Self-reported table.Published scores are OpenAI's, not LLM Stats verified. GPQA and ARC are near-ceiling.
• Gateway default effort is low. OpenAI markets Highest reasoning, so set the effort explicitly when that matters.
• Critical cyber designation. Most advanced cyber capabilities start limited for trusted defenders.
• Knowledge cutoff April 30, 2026.
• No fine-tuning on this card.
• Path to Astra states Astra was not involved in the Hugging Face incident.

Outlook

This post ages when LLM Stats verifies Terminal-Bench 4.0, DeepSWE, or OSWorld, or when OpenAI changes rollout coverage or pricing. Through-line: released GPT-6 flagship with a self-reported benchmark table, a $10 / $50 Standard sticker, long-context multipliers, and gateway default effort low.

Sources: OpenAI GPT-6 Astra announcement, model docs, Path to Astra.